Synthetic Dataset for Causal Analysis with Localized Latent Variables and Nonlinear Dependencies

This dataset provides a synthetic benchmark for causal discovery in the presence of localized latent variables and nonlinear dependencies. The data are generated from structural causal models (SCMs) consisting of multiple subgraphs, where latent variables influence only specific subsets of observed variables. The dataset includes nonlinear relationships, confounding structures, and missing data mechanisms. The dataset contains:- observed variables- latent ground truth variables (for evaluation purposes)- the true causal graph structure- code for data generation and reproducibility- scripts for causal effect estimation (ATE) and graph evaluation (SHD, precision, recall) A Missing At Random (MAR) mechanism is introduced in part of the data, allowing evaluation under incomplete observations. This dataset is intended for:- benchmarking causal discovery algorithms- evaluating latent variable reconstruction methods- studying the impact of confounding on causal effect estimation All data are fully reproducible via the provided Python scripts.